Robust RL-based map-less local planning: Using 2D point clouds as observations

Abstract

In this letter, we propose a robust approach to train map-less navigation policies that rely on variable size 2D point clouds, using Deep Reinforcement Learning (Deep RL). The navigation policies are trained in simulations using the DDPG algorithm. Through experimental evaluations in simulated and real-world environments, we showcase the benefits of our approach when compared to more classical RL-based formulations: better performance, the possibility to interchange sensors at deployment time, and to easily augment the environment observability through sensor preprocessing and/or sensor fusion. Videos showing trajectories traversed by agents trained with the proposed approach can be found in https://youtu.be/AzvRJyN6rwQ.